SUO: RE: Enhancing Data Interoperability with Ontologies...
John F. Sowa wrote:
> John,
>
> That's the critical point:
>
> JB> There are some obvious reasons for this:
> > anaphora resolution relies heavily on real-world
> > knowledge and that is not available in great
> > quantities for the anaphora resolution
> > algorithms. The standard techniques of picking
> > the most likely referents work pretty well (in
> > fact very well), but there is always that small
> > residue on top that they currently miss.
>
> Natural languages have evolved for use by
> people who apply both syntactic and semantic
> knowledge to interpret a text. Some NL
> processors have achieved a fairly good level
> of accuracy by using syntax alone, and some
> have achieved a fairly good level by using
> mostly semantics with a modest amount of syntax.
> But in general, both are needed.
>
> Even so, no NL processor (human or computer)
> can ever achieve 100% reliability on every text
> thrown at it because nobody (human or computer)
> is omniscient. And even if the reader is
> very well versed in the subject matter, there
> is no guarantee that the author has written
> crystal-clear prose that eliminates all
> conceivable misinterpretations (as we have
> all seen in numerous email exchanges).
>
> Therefore, no NL processor can ever achieve
> 100% reliability. However, it is also important
> to realize that the supposedly unambiguous
> artificial languages (logic, controlled NLs,
> or any programming language ever invented)
> are just as unreliable -- but the authors
> are usually blamed rather than the readers.
>
> Fundamental principle: A formally precise
> language can give no assurance of reliability
> because what it says so precisely may have no
> relationship to what the author intended.
>
> John Sowa
So a wide comprehension NL question answering
system isn't on the horizon, and would take
an enormous investment to build up a knowledge
base of world knowledge.
But a more narrowly focused NL question
answering system could in principle ask
questions of the user wherever it "realizes"
that its knowledge is incomplete. But
with anaphora resolution of some kind, it
would appear more natural than without
anaphora.
Suppose there is such a program. Could
it use the essential principles of little
world knowldege, but with anaphora, and
using DRT, record the session with a
user and, after completing the session,
record any knowledge it has acquired
as a more elemental CLCE text?
That approach might provide the precision
that is ultimately needed without the
stilted language putting off the users.
Is that approach feasible?
Thanks,
Rich